Tackling System and Statistical Heterogeneity for Federated Learning with Adaptive Client Sampling

被引:123
作者
Luo, Bing [1 ,2 ,4 ,5 ]
Xiao, Wenli [1 ,2 ]
Wang, Shiqiang [3 ]
Huang, Jianwei [1 ,2 ]
Tassiulas, Leandros [4 ,5 ]
机构
[1] Shenzhen Inst Artificial Intelligence & Robot Soc, Shenzhen, Peoples R China
[2] Chinese Univ Hong Kong, Sch Sci & Engn, Shenzhen, Peoples R China
[3] IBM Corp, TJ Watson Res Ctr, Yorktown Hts, NY USA
[4] Yale Univ, Dept Elect Engn, New Haven, CT 06520 USA
[5] Yale Univ, Inst Network Sci, New Haven, CT 06520 USA
来源
IEEE CONFERENCE ON COMPUTER COMMUNICATIONS (IEEE INFOCOM 2022) | 2022年
关键词
OPTIMIZATION;
D O I
10.1109/INFOCOM48880.2022.9796935
中图分类号
TP3 [计算技术、计算机技术];
学科分类号
0812 ;
摘要
Federated learning (FL) algorithms usually sample a fraction of clients in each round (partial participation) when the number of participants is large and the server's communication bandwidth is limited. Recent works on the convergence analysis of FL have focused on unbiased client sampling, e.g., sampling uniformly at random, which suffers from slow wall-clock time for convergence due to high degrees of system heterogeneity and statistical heterogeneity. This paper aims to design an adaptive client sampling algorithm that tackles both system and statistical heterogeneity to minimize the wall-clock convergence time. We obtain a new tractable convergence bound for FL algorithms with arbitrary client sampling probabilities. Based on the bound, we analytically establish the relationship between the total learning time and sampling probabilities, which results in a non-convex optimization problem for training time minimization. We design an efficient algorithm for learning the unknown parameters in the convergence bound and develop a low-complexity algorithm to approximately solve the non-convex problem. Experimental results from both hardware prototype and simulation demonstrate that our proposed sampling scheme significantly reduces the convergence time compared to several baseline sampling schemes. Notably, our scheme in hardware prototype spends 73% less time than the uniform sampling baseline for reaching the same target loss.
引用
收藏
页码:1739 / 1748
页数:10
相关论文
共 44 条
[1]  
Alain G, 2016, Arxiv, DOI arXiv:1511.06481
[2]  
Avent B, 2017, PROCEEDINGS OF THE 26TH USENIX SECURITY SYMPOSIUM (USENIX SECURITY '17), P747
[3]  
Bonawitz K., 2016, NIPS WORKSHOP PRIVAT
[4]  
Bonawitz K., 2019, ARXIV, P374
[5]  
Boyd S., 2004, Convex optimization
[6]   Convergence Time Optimization for Federated Learning Over Wireless Networks [J].
Chen, Mingzhe ;
Poor, H. Vincent ;
Saad, Walid ;
Cui, Shuguang .
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, 2021, 20 (04) :2457-2471
[7]  
Chen WL, 2021, Arxiv, DOI arXiv:2010.13723
[8]  
Gopal S, 2016, PR MACH LEARN RES, V48
[9]  
Haddadpour F, 2019, Arxiv, DOI arXiv:1910.14425
[10]  
Cho YJ, 2020, Arxiv, DOI arXiv:2010.01243